MARS internal memo · 2026 Q3

State of Play.
Where we are, honestly.

A hundred-and-eighty degree look at the substrate, how it compares to CB and PB, what the platform-verticals pattern is unlocking, and what a $0-marginal-cost growth path actually looks like.

To · Adrian, Raul, futures, proxy, rss, reit — the mars team, in the useful sense of the word

01 The substrate today

What we actually have

Every number below is a live count. If any of it surprises you, that's worth a conversation on its own.

Companies
251,677
186,595 CIK-linked · 65,082 news/spider-only
Persons
1.20 M
453K bios · 21K work-history · 20K education
Deals · funding + M&A
113,807
$60.7 T tracked in verified/vouched tier
SEC filings
486,177
Form D · 144 · 13D · 13G, all direct-to-v2
This is a publicly-invisible version of a company graph. We built it by joining news-extracted deals to SEC filings to real-estate transactions to person filings — all on one canonical company_id.

What sits inside

Deal substrate

77,662 funding deals (Layer 2 canonical) sitting on top of ~140K article-level substrate rows. 36,145 canonical mergers. 5,679 real-estate deals as of two days ago and growing. 107K bridge rows tying every deal to the articles that produced it — verifiable line by line.

Entity substrate

99,909 investors — of which 51,247 are ADV-enriched with CRD, RAUM, disciplinary flags. 5,818 advisors. 5,150 institutions (education, employers). 1.2 M canonical persons resolved from Form D, Form 144, news extractions, and — increasingly — SEC insider trading paths.

Who did what got us here

adrian
SEC forms pipeline. Form D, Schedule 13D/G, Form 144 — 486K filings in v2, most of what makes the substrate defensible. Six months ago we had zero of these.
raul
M&A canonicalization. The mergers pipeline that runs upstream of our Layer 2 canonical merger deals; the porting of his industry taxonomy + COMPANY_ALIASES + SHA256 hash key gave us reproducibility with the mergers_grouped canonical.
rss
Premium press coverage. Bloomberg / WSJ / FT paths that GDELT structurally under-indexes. Fills the exact hole in our news layer.
proxy
Spider bandwidth economics. Asset blocking + tiered routing work; ~30% bandwidth reduction shipped 2026-07-02. This is what keeps unit costs sane as coverage grows.
reit
First vertical. Scaffolded 2026-07-02, shipping 5,679 real-estate deals by 2026-07-04 — following every platform contract, zero mars-side code changes. This validates the pattern for every other vertical we'd spin up next.
futures
Consumer. Reads the substrate through MCP tools. The pipe_deals_v materialized view + Atlas ticker join went live for their consumption 2026-07-02.

02 Coverage geography

Where the substrate is deep and where it's thin

Every country below is a real count of companies with at least one deal or filing.

Tier Regions Active companies What that unlocks
Flagship United States 15,017 Full editorial coverage, weekly cadence, everything
Deep Canada 2,739 Weekly cadence today, no additional ingest work
Ready for
monthly V1		 United Kingdom 585 One SQL query + template + localization pass away. Zero new ETL.
Singapore249
China / HK245
Israel229
France186
Germany164
India · Australia · Japan · Netherlands · Sweden96–155 each
Quarterly
cadence		 Mexico 57 Enough for a real recap. Coverage grows monthly through GDELT + Reg S filings.
Spain52
Brazil51
Ireland · Italy38–49
Needs work Rest of LatAm · SE Asia · Africa · MENA <20 each Ingest gap. Sources exist; we just haven't added them yet.

The interesting split is between "ready for a Vol. 1 tomorrow" (row 3) and "quarterly for now" (row 4). Both of those tiers are already viable products; nobody needs to build anything for us to publish them.

03 vs Crunchbase, vs PitchBook

Where we compete, where we don't, and why it matters

Honest read. Chest-thumping doesn't help us decide what to build next.

Capability MARS CB PB Notes
Basic company graph (US + tier-1 intl) ● ● ● Roughly at parity on covered companies
Direct SEC ingest (Form D / 144 / 13D/G / ADV) ● ◐ ◐ Both source it; neither surfaces it as first-class data
Canonical entity layer across news + filings + RE ● ○ ◐ One company_id across everything is our structural difference
Real-estate deal substrate ● ○ ◐ reit vertical went live 2 days ago; PB has it via a separate product
Deep LP data (commitments, IRRs) ○ ○ ● PB moat. Structural — comes from LPX + years of surveys
Fund performance benchmarks ○ ○ ● Preqin / Cambridge / ILPA territory — buy, don't build
Deep private-credit substrate ◐ ○ ● Credit vertical scaffolded; SEC substrate covers BDCs today
Cross-linking to Form 144 insider activity ● ○ ○ CoreWeave workup demonstrated this — pre-IPO officers = post-IPO sellers
MCP tool set / programmatic access ● ◐ ● Ours ships direct-to-substrate; theirs adds a middle layer
Brand recognition · sales channel ○ ● ● This is the actual gap. Not technical.

Read: we're structurally at parity on covered companies, ahead on cross-linking / SEC substrate / canonical entity architecture, behind on LP data + fund performance + brand recognition. The last one is the real gap — technical parity is not customers. The verticals + newsletter format are the wedge for that.

04 The verticals architecture

Rails, not a monolith

Codified 2026-07-02 in MARS_PLATFORM_CONTRACTS.md. reit tested it. It works.

The strategic call this quarter: MARS becomes the shared rails, verticals become their own parallel projects. Every vertical writes direct to v2 under contract:

  1. FK to companies_v2, no parallel company table
  2. Standard integrity_tier pattern
  3. Own methodology prefix (re-*, credit-*, ...) recognized by the health check
  4. Own mars.<vertical>_deals_v2 table
  5. Ships direct-to-v2, no v1 detour
  6. Deterministic + LLM + audit trio for every extractor
  7. Health-email visibility via check_<vertical>
reit proved it: scaffolded empty-project on 2026-07-02, shipping 5,679 real-estate deals by 2026-07-04. Zero mars-side code changes to accommodate them. The merge_entities tool picked up their FK automatically; the health check recognized their methodology prefix without config. That's what "rails" means.

What's in flight

reit · live

5,679 deals. 4,192 distinct companies. 40% Form-D-linked. First health check found real issues on day 1 ($5T Nuveen extraction, unit-drop rows) — reit team fixing at source. Product path: monthly real-estate recap for CRE investors, LP allocators.

private credit · scaffolded

Project skeleton at /home/kee/code/private_credit/, waiting for owner. Sources documented (BDC 10-Q Schedule of Investments, Form D pooled funds, ADV, N-CSR/N-PORT). Product path: monthly private-credit landscape recap. Zero external data purchase required.

Europe / international · substrate exists

Coverage table above shows 585 UK, 186 France, 164 Germany, 155 Australia, 111 India — all with recent deal activity. What's missing is the editorial layer, not the substrate. Same shape as Mexico Ticker (shipped 2026-07-04) but scale up.

Others in queue

Insurance products, biotech clinical-stage tail, LP-data (buy don't build). Each spun up the same way — as a separate project with its own owner, plugging into rails via the contract.

05 The cost story

Local models change the growth math

This is the section that changes what the roadmap can look like.

The default assumption in this space is that every additional edition, vertical, or language means additional FTE + additional API spend. Neither is true for us at current unit economics.

The tier chain in numbers

Every entity-resolution or classification decision routes through:

Tier Model Where it runs Cost / call Handles
ADeterministic rulesPython + Postgres$0Strong-ID matches, sentinel filter, most of the volume
BQwen3-30B via Infernobench3 (our GPU)$0Bulk classification, SAME/DIFFERENT pair verdicts
CHaiku 4.5Anthropic API~$0.001Escalated ambiguity, structured extraction
DGrok-4 + web_searchxAI API~$0.08Real-world knowledge — final tier only
Real numbers: 12,697 pending entity-resolution rows cleared this week for ~$35. All-Grok projection was ~$1,013. That's 29× cheaper, not a rough estimate — measured on the actual run.

What that unlocks

Every edition costs pennies

Producing the US Ticker: one SQL query, one template, a couple of Qwen classification calls. Producing the Mexico Ticker in Spanish: identical cost. The English → Spanish translation is a Qwen call away and costs nothing at bench3 rates. Same for a hypothetical German edition, or a Portuguese Brazil edition.

Every vertical carries its own budget

Per the platform contracts, each vertical brings its own external AI + proxy spend. mars operates shared infra but doesn't subsidize. So as verticals multiply, MARS platform cost stays flat.

Every audit sweep is free

Weekly re-audit of the 5-year clean_domains substrate through the tier chain: still fits in the $0-cost tier for the deterministic gate + Qwen batch. Anthropic + xAI only kick in on residue.

Growth without OpEx cliff

The traditional cost curve — headcount scales with coverage — doesn't apply to us. What scales is engineering effort per vertical (bounded, one owner), and shared infra usage (linear in queries, not in coverage). This is why the verticals architecture is the actual moat.

06 Concrete next moves

If we agree with the read above, what do we do about it?

Six proposed moves. Each is a specific team-level ask; each can start or defer without dependency chains.

01
Ship the Q3 editions calendar. Weekly US Ticker starting this Sunday. Monthly UK, Singapore, Israel, France, Germany starting late July. Quarterly Mexico, Brazil, Netherlands, Sweden starting August. All from the existing substrate — no new ingest work.
Owner: mars / kee. Cost: pennies per edition.
02
Find owners for private credit. The scaffold is at /home/kee/code/private_credit/. The sources are documented. What we need is a person with a private-credit background to own the extractor + editorial voice. Bring names, or defer.
Owner: TBD. Cost: one FTE-equivalent, once found.
03
Adrian: SEC substrate freshness push. Form 4 (insider transactions) is the next tier of value alongside the existing Form 144. Adding it lets the CoreWeave-style "pre-IPO officer → post-IPO seller" story extend to every ownership transition. Estimated: same shape as Form D pipeline, ~1 week.
04
Raul: M&A advisor coverage push. merger_advisors_v2 shipped 2026-06-28 with 31K rows and 27% raw coverage. Getting to 50% coverage unlocks the "who's on which side of every deal" surface that Bloomberg charges $30K/seat for. Estimated: extractor enhancement + backfill, ~2 weeks.
05
rss: premium-press latency check. Bloomberg/WSJ/FT ingest via rss vertical is the direct answer to GDELT's premium-press gap. Where are we on landing that in a form MARS consumers can read against a canonical docid?
Owner: rss. Ask: status + expected landing quarter.
06
proxy: cost-per-1000-crawl target. With the 2026-07-02 asset-blocking work banked (~30% reduction), what's the next meaningful lever? Tiered routing? DC-tier for public-source cache-hits? Set a Q3 target and hit it.
Owner: proxy. Ask: numeric target + one-line plan.

How this was built. Every count above is a live SQL query against the v2 substrate. The comparison table (section 03) is directional; specific CB / PB coverage numbers move around and I didn't want to cite anything I couldn't verify against a recent source. Read those rows as trend, not benchmark.

Companion pieces. Three-company workup shows the substrate at per-company depth. US Ticker Vol. 1 is the current-events companion. Mexico edition demonstrates the localization pattern that Section 05 argues we can replicate at negligible cost.

What this document isn't. A commitment. A pitch. A roadmap. It's the state of play as of 2026-07-04 evening, published so we can agree or disagree with the read before the next planning cycle.

MARS · 2026-07-04 · internal team memo · kee@kscope.io